Physical reasoning about humans in video has mostly centered on everyday interactions—contact, pressure, pose, and motion—where these cues support embodied perception and contact-aware applications. But short, forceful events, such as impacts, have received less attention. A new arXiv paper introduces HAPACT, a benchmark designed to fill that gap by localizing physical impacts in movie scenes.
The benchmark specifically targets human-centric physical impact localization, meaning it aims to identify where and when forceful interactions occur. Because the abstract is truncated, full details on the dataset's construction and evaluation metrics are not yet available, but the stated motivation is clear: to push video reasoning beyond gentle, sustained contact toward abrupt and forceful events.
By focusing on movies, HAPACT offers a challenging testbed with varied cinematography, occlusions, and realistic action. The authors position this as a step toward more robust embodied perception, where recognizing impacts could benefit applications in activity recognition, robotics, and content analysis.